A premortem assumes your plan already failed six months from now, then works backward to find every reason why. The technique was published by Gary Klein in Harvard Business Review. Daniel Kahneman called it his single most valuable decision-making technique. Google, Goldman Sachs, and Procter & Gamble run them before major decisions.
This repo packages the premortem as a portable AI skill. Install it once, then any time you have a plan you care about, ask Claude (or Codex, or any agent that loads skills) to premortem it. You get back a short, scannable breakdown right in the chat: how the plan died, the assumptions that made the failure possible, and a revised plan you can act on this week.
For Claude Code:
git clone --depth 1 https://github.com/PabloNAX/premortem-skill && ./premortem-skill/install.sh claudeFor Codex or generic agents, swap the target: ./premortem-skill/install.sh codex or ./premortem-skill/install.sh agents. Use all to install everywhere.
Then ask Claude (or Codex) to premortem this: <your plan> and the skill takes over. The full install reference is further down.
Example prompts that trigger the skill:
premortem this launch: I am running a $297 workshop for 50 marketing managers next month
premortem my pricing change from $19 to $49
what could kill this hire?
stress test this strategy pivot to enterprise
am I missing anything before I ship this?
poke holes in this product plan
The skill replies in chat with a short, scannable premortem: most likely failure, most dangerous failure, the hidden assumption, a revised plan, and a pre-launch checklist.
Ask Claude "is this a good plan?" and it finds reasons to say yes. Default behavior is agreement. The premortem flips the frame. You tell Claude the plan is already dead, and Claude stops looking for reasons it works and starts explaining how it fell apart. Researchers at Wharton and Cornell named this "prospective hindsight" and showed it sharpens failure prediction by a wide margin.
The reframe is the whole product. Everything else is structure.
A short, scannable premortem written straight into the chat. No files, no HTML, no build step.
The synthesis sits at the top: the most likely failure, the most dangerous failure, and the hidden assumption you never questioned. Below it, one line per failure mode with an early-warning signal to watch for, then a revised plan and a short pre-launch checklist. Every item is concrete enough to act on this week.
It stays in the conversation, so you can push back, ask for more depth on one failure, or re-run it as the plan changes. Want a copy? Ask, and it saves the same breakdown as a single Markdown file.
- Gathers context. The skill checks the conversation,
CLAUDE.md, memory folders, and attached files. If it can describe the plan, the audience, and what success looks like, it proceeds. If a piece is missing, it asks one question at a time until the threshold is met. - Sets the frame: "It is six months from now. This plan has failed. We are looking back to understand why."
- Generates failure reasons. One pass, no fixed count. Four for simple plans, nine for complex ones. Every reason is specific to your plan, not boilerplate.
- Spawns sub-agents in parallel, one per failure reason. Each agent writes the failure story, names the underlying assumption, and gives early warning signs you can observe.
- Synthesizes everything into one compact breakdown and writes it into the chat: synthesis, failure modes, revised plan, and a pre-launch checklist.
If you already cloned the repo, the installer takes a target:
./install.sh # all targets
./install.sh claude # Claude Code only
./install.sh codex # Codex only
./install.sh agents # generic AGENTS_HOME
./install.sh --dry-run # preview without writingDefault install paths:
Codex: ${CODEX_HOME:-$HOME/.codex}/skills/premortem
Claude: $HOME/.claude/skills/premortem
Agents: ${AGENTS_HOME:-$HOME/.agents}/skills/premortem
For Claude.ai web: zip skills/premortem/ and upload it under custom skills.
For any other agent or editor: copy skills/premortem/ into that tool's skills/prompts/rules directory, then tell it Use premortem to stress-test my plan.
Run it when:
- You are about to launch something with money or reputation on the line.
- You are making a hire.
- You are changing pricing or business model.
- You are pivoting strategy or positioning.
- You are committing to a deal or partnership.
- You are shipping a feature where rollback would be expensive.
Skip it when:
- You have a vague idea, not a concrete plan. Plan first, premortem second.
- The question has one right answer. Just answer it.
- You want feedback on a draft. That is editing.
- The decision is already made and locked. A premortem only helps when you can still change course.
The two questions sound similar. They run on different psychology.
"What could go wrong?" produces cautious lists. People hedge. They mix likely problems with unlikely ones because both feel safer to mention than to omit.
"This already failed, explain why" produces stories. The brain shifts into post-hoc reasoning mode and constructs specific causal chains. The output is concrete enough to act on. Kahneman points to this as the rare debiasing technique that consistently works.
This skill encodes the prompt that triggers that second mode, plus the structure that turns the output into a breakdown you can hand to a co-founder or a team.
The installable skill lives here:
skills/premortem/
SKILL.md # workflow instructions
agents/
openai.yaml # OpenAI agent manifest
Do not upload the repo root as the skill folder. Use skills/premortem/.
The root SKILL.md is a working copy for local development. Keep it in sync with skills/premortem/SKILL.md before publishing.
None. The skill is plain Markdown plus a small install script. No Python, no Node, no npm, no server.
| Surface | Status | Install path |
|---|---|---|
| Codex | supported | $HOME/.codex/skills/premortem |
| Claude Code | supported | $HOME/.claude/skills/premortem |
| Claude.ai | supported | upload zipped skills/premortem/ |
| Generic agent | supported | $HOME/.agents/skills/premortem |
| Other editors | manual | copy skills/premortem/ into prompts |
bash -n install.sh
./install.sh --dry-runThe premortem technique was published by Gary Klein in Harvard Business Review (2007). Daniel Kahneman has cited it as his single most valuable decision-making technique. Research on prospective hindsight comes from work by Mitchell, Russo, and Pennington at Wharton and Cornell.
MIT. See LICENSE.